A Machine Learning Approach for Scalable Early-Age Dyslexia Detection Correlating Phonological Speech Analysis and Eyeball Tracking - On My Own ...

Page created by Raymond Hughes
 
CONTINUE READING
A Machine Learning Approach for Scalable Early-Age Dyslexia Detection Correlating Phonological Speech Analysis and Eyeball Tracking - On My Own ...
iJournals: International Journal of Software & Hardware Research in Engineering (IJSHRE)
 ISSN-2347-4890
 Volume 9 Issue 7 July 2021

A Machine Learning Approach for Scalable
 Early-Age Dyslexia Detection Correlating
 Phonological Speech Analysis and Eyeball
 Tracking
 First Author: Kareena Shankta - Dhirubhai Ambani International School
 Second Author - Reetu Jain Chief Mentor On My Own Technology Pvt Ltd
 Third Author – Shekhar Jain CEO On My Own Technology Pvt Ltd
 Fourth Author – Radhakrishna Midas - Sr. Research Associate at On My Own Technology

Abstract – Dyslexia concerns a persistent and unexpected difficulty in developing age and experience
appropriate word reading skills, which encompasses proficiency in accuracy and efficiency. It is termed a
reading disability that at an early age is extremely hard to trace because the degree and type of dyslexia are
unique to an individual, their genetics and environmental factors. While there is yet no full understanding of the
cause of dyslexia, or agreement on its precise definition, it is certain that many individuals suffer persistent
problems in learning to read for no apparent reason. Since this could drastically affect literacy skills and
academic engagement, there is a need to address the quantifiable screening process for dyslexia as early,
effectively, and objectively as possible.

Our solution attempts to screen dyslexia problems for children at every age. This mechanism correlates an
eyeball tracking and speech-recognition technology, written in python with Matplotlib as the main plotting
library. The solution utilised is a scalable measure of dyslexia at every age. While the user is asked to read a
passage on the screen, the camera traces the pupil’s positions with respect to eye momentum and direction,
recording the independent variables- fixations and saccades. By analyzing the distribution pattern of existing
datasets, the model makes a prediction. The user is then asked to read aloud a series of 25 words, while the
algorithm processes the speech and identifies any discernible differences in speech-features (associated with
dyslexia). The independent variables (input parameters) are reaction time, backtracking, unexpected
pronunciation and reading time. The final algorithm prediction provides results on the predicted level of
dyslexia.

This novel solution aims to give estimative predictions by combining eyeball tracking (reading) and speech
recognition (phonology) so that the early detection of dyslexia leads to early intervention.

Keywords – dyslexia analysis, eyeball tracking, fixations, saccades, phonological testing.

Kareena Shankta; Reetu ,jain; Shekhar Jain; Radhakrishna Midas, Volume 9 Issue 7, pp 1-10 July 2021
A Machine Learning Approach for Scalable Early-Age Dyslexia Detection Correlating Phonological Speech Analysis and Eyeball Tracking - On My Own ...
iJournals: International Journal of Software & Hardware Research in Engineering (IJSHRE)
 ISSN-2347-4890
 Volume 9 Issue 7 July 2021

 1. INTRODUCTION the disability of dyslexia, especially in young
 Dyslexia concerns a persistent and unexpected children, given that early intervention goes a long
 difficulty in developing age and experience way in paving high self-confidence for children
 appropriate word reading skills, which while yielding a high literacy-rate for the country.
 encompasses proficiency in accuracy and
 1
 efficiency. It is the most studied and the most Underlying skill reading is a multifaceted network
 known learning disorder in the world, yet is still of competencies for both understanding language
 one of the hardest to diagnose and resolve. meaning and recognizing printed words. General
 According to Geurrero et al,, it can have different language competencies include knowledge of word
 clinical manifestations in different phases of life meaning, facts and concepts as well as mastery of
 and can influence different dominions of life. In grammar and the structure of language at the
 2019, 131 million students enrolled in government sentence, text and discourse levels. In accordance
 schools, joining the estimated 747 million children with the ideology of the Guerrero et al paper, we
 of primary-school age worldwide (United Nations will be considering individuals typically displaying
 Department of Economic and Social Affairs, a broad spectrum of ability in both reading and its
 20192), where they embarked on their journey of precursor skills as lower end of the spectrum
 learning to read and write, skills that will carry termed in clinical literature, “specific learning
 them through high school and beyond. For up to disability in reading” or “developmental dyslexia”.
 20% of these children across different languages
 and countries, however, this journey carries with it This broad-spectrum inclusion supersets dyslexia.
 a risk of school failure due to an inability to learn Therefore, for the sake of clinical accuracy, the
 to read as expected. A report released by the Indian term dyslexia and reading disabilities will be used
 government in 2015 estimated that at least 35 interchangeably. International Dyslexia
 million Indian children are dyslexic with millions Association recognizes the term Evaluation as a
 more suspected to be undiagnosed. Intervention more accurate word to describe the process of
 research indicates that targeted, evidence-based determining if someone has dyslexia. The
 programs in kindergarten and first grade can word evaluation encompasses identification,
 substantially improve reading outcomes for screening, testing, diagnosis, and all the other
 children who require intensive intervention in order information gathering involved when the student,
 to become successful readers. However, pre- and his or her family, and a team of professionals work
 beginning readers are not routinely screened for together to determine why the student is having
 risk in part due to the formidable logistical and difficulty and what can be done to help. Evaluation
 financial challenges of providing millions of conventionally involves a three-stepped process of
 children with a full clinical assessment of the information-collection from parents and
 language and cognitive skills that are precursors to educationalists, screening and testing; and
 reading. There exists a need to objectively identify intervention plan leading to conclusion and
 recommendations.
 1

 https://dictionary.apa.org/dyslexia#:~:text=n.,words%20and%20
 sometimes%20in%20arithmetic.
 2 https://unstats.un.org/sdgs/report/2019/The-Sustainable-
 Development-Goals-Report-2019.pdf

 2
© 2021, iJournals All Rights Reserved www.ijournals.in Page 2

© 2020, iJournals All Rights Reserved www.ijournals.in
A Machine Learning Approach for Scalable Early-Age Dyslexia Detection Correlating Phonological Speech Analysis and Eyeball Tracking - On My Own ...
iJournals: International Journal of Software & Hardware Research in Engineering (IJSHRE)
 ISSN-2347-4890
 Volume 9 Issue 7 July 2021

 CURRENT DIAGNOSTIC CRITERIA standardized tests and the documented history of
 learning difficulties, most of all for individuals
 The American Psychiatric Association has created aged 17 and older, where the use of standardized
 the international diagnostic criteria for learning tests could not be valid, sensible and specific
 disabilities in DSM-5. The diagnostic criteria used enough to describe the quality and the quantity of
 for specific learning disorders are four: the learning disorders [16]. Some Years ago,
 Criterion a) the persistency of the symptoms for at UtaFrith, one of the most known scientists in the
 least 6 months, despite specific interventions; field, stated “diagnosis based on behavioural
 Criterion b) the impairment of single or more criteria will always be problematic, as behavior is
 abilities, with negative effects on school strongly influenced by comorbidity, motivation,
 achievement; age, and ability” and “Adults with dyslexia and
 adults with autism who show good compensation
 The second diagnostic criteria refers to the specific can sail through standard diagnostic tests. Have
 affected academic skill (one or more than one), that they ceased to be dyslexic or autistic? Of course
 must be substantially and quantifiably below those not. They themselves tell us about their persisting
 expected, based on the chronological age, and it problems, and we can reveal these problems with
 negatively influences academic achievement, or more subtle tests.
 occupational performance, or activities of daily
 life. With reference to the impairment in a single or criterion d) last but not least, there are different
 more abilities, from a psychometric point of view exclusion criteria.
 the cut-off that is to be used is one and a half The first exclusion criterion refers to absence of
 deviation standard below the expected mean in intellectual disability.
 each single evaluated ability (or even one standard The second refers to the exclusion of a fragmentary
 deviation in some cases, or under the 5th percentile and inadequate instruction.
 for some measure) [18]. The third refers to linguistic proficiency in the
 Criterion c) the onset in a school-age, even if the language used for academic instruction.
 disorder could fully manifest later; The fourth refers to absence of sensory problems
 high enough to justify the learning difficulties
 With reference to the age of onset, there is an (visual and auditory sensory problems) [16,17,19].
 agreed awareness that the clinical manifestations of
 learning disorders can change during the
 2. LITERATURE REVIEW
 individual’s life and that they are a lifelong
 condition [6]. Generally they are discovered in the Current diagnostic tools:
 first years of schooling, but it is also possible that
 they could be discovered later in the individual’s In India, the standardized procedure involves
 schooling history. In order to take into account the DALI, tools for school teachers and psychologists
 clinical manifestations in different phases of life, in Indian Languages to identify dyslexia. It works
 there is a clear reference to the schooling history by asking a series of questions related to reading
 and the clinical history, to the integrated use of and word decoding abilities. DALI as an

 3
© 2021, iJournals All Rights Reserved www.ijournals.in Page 3

© 2020, iJournals All Rights Reserved www.ijournals.in
iJournals: International Journal of Software & Hardware Research in Engineering (IJSHRE)
 ISSN-2347-4890
 Volume 9 Issue 7 July 2021

 assessment tool was created by the National Brain data-points to determine to what extent is a person
 Research Centre, India in South Asian languages at risk of developing dyslexia.
 with the primary aim of acting as a bridge for
 dyslexia diagnosis in a multilingual country of A study by Benfatto et al indicates, “Eye tracking
 India to identify dyslexia in indigenous languages can be an efficient means to identify children at
 like Hindi, Marathi, Kannada and English. Despite risk of long-term reading difficulties' '. The method
 its dire necessity, it is very subjective and relies on involved originates from KronobergsProjektet, a
 a child’s/mentor’s interpretation of his/her own research project on reading development and
 symptoms, which can make children self-conscious reading disability in Swedish school children. The
 of their “problem”. the systems using a computer- project ran between 1989 and 2010. Eye-tracking
 based battery of tests or questionnaires without during reading was conducted on children in 3rd
 introducing a new algorithm on top. They are just grade, and reading difficulties were assessed until
 the digital transcription of human expert tests, and adulthood. What makes the KronobergsProjektet
 predictions are usually based on statistics. study unique is: 1) children were monitored over a
 long period of time and 2) the integrity of the
 Several studies have been conducted internationally recorded eye movement data.
 in the hopes of diagnosing dyslexia with the help of
 Artificial Intelligence, a much awaited catalysis for Methods and Results: The experiments are based
 this topic of research in the neuroscience on eye tracking data from a cohort of 185 subjects
 community. Currently, validated diagnostics are participating in the project. Each participant (all
 delivered by accredited professionals based on Swedish speakers) has to read one and the same
 scoring of a diverse battery of tests. Artificial text presented on a single page of white paper with
 Intelligence and Machine learning have been high contrast. For each session, using eye-trackers,
 utilised to provide accurate results at a faster rate the authors record a lot of numerical parameters
 and easier convenience. among which, for instance, the duration of the
 event, the distance spanning the event, the average
 eye position during the event, the maximum range
 EYE_BALL TRACKING between any two positions, etc. Having gathered all
 these data and properly preprocessed their data set,
 Eye-tracking is a process more recently used in the authors have implemented a standard classifier
 Artificial Intelligence as a measure of either the showing a high degree of accuracy (around 96%).
 point of gaze or the motion of an eye relative to the It is important to note that the model also takes into
 head. An eye tracker is a device for measuring eye account the results from a battery of other common
 positions and eye movement. Previous research tests, such as rapid automatized naming, reading of
 (xxx) has indicated that dyslexia may be correlated non-words, etc. One interesting feature of the
 to a dysfunction in the oculomotor process for process is that the prediction can occur after as
 which two main independent variables of eye little as 30 seconds of reading for the
 fixations and saccades were sourced as information user.According to the study, “Using statistical
 cross-validation techniques on a sample of 97 high-

 4
© 2021, iJournals All Rights Reserved www.ijournals.in Page 4

© 2020, iJournals All Rights Reserved www.ijournals.in
iJournals: International Journal of Software & Hardware Research in Engineering (IJSHRE)
 ISSN-2347-4890
 Volume 9 Issue 7 July 2021

 risk and 88 low-risk control subjects, they achieved signal processing can extract patterns that are not
 a classification accuracy of 96%.” Thus, in this accessible to a human expert. Hence, we decided to
 study we will be using these as our literature values use a factor of both eye-ball tracking and
 for comparison. phonology testing as variable markers for dyslexia.

 Necessary definitions: A saccade is a quick We look for backtracking,
 simultaneous movement of both eyes between two
 or more phases of fixation in the same direction. Based on this research, we hypothesized a
 phonological speech analysis solution, which could
 According to Castles et al, from the phonological be correlated with eyeball tracking to improve
 deficit hypothesis, several parameters can be accuracy.
 derived. A study conducted by lead researchers at
 Dystech focused on utilising audio recordings of
 participants reading out words as a method of
 discerning dyslexic readers from non-dyslexic
 readers. For every audio record, they considered 2
 parameters: Figure 1: Parameters for the Project

 ● The Reading Reaction Time (RRT) which
 It is important to move from a deficit to a
 is the interval between the initial display of the
 preventive model of education.
 word and the start of the reading.
 ● The Reading Time (RT) which is the time
 it takes for the participant to read the corresponding 3. DYSLEXIA
 word.
 ● RRT and RT evaluation is done via a Dyslexia is not caused by poverty, developmental
 computer (no human in the loop). Consequently, delay, speech or hearing impairments, or learning a
 from a session of 32 audio records, we extract 6 second language, although those conditions may
 numbers which will be used in our ML put a child more at risk for developing a reading
 experiments: disability (Snow, Burns, & Griffin, 1998).

 ● average for 32 words, average 
 Children with dyslexia will often show two obvious
 for 16 real words, average for 16 nonsense
 difficulties when asked to read text at their grade
 words
 level. First, they will not be able to read as many of
 ● average RT for 32 words, average RT for
 the words in a text by sight as average readers.
 16 real words, average RT for 16 nonsense words
 There will be many words on which they stumble,
 guess at, or attempt to "sound out." This is the
 It is to be noted that poor reading performance is
 problem with "fluent word recognition" identified
 not an ultimate marker of dyslexia, but Dystech
 in the previous definition.
 results demonstrate that a dedicated machine
 learning algorithm associated with proper audio

 5
© 2021, iJournals All Rights Reserved www.ijournals.in Page 5

© 2020, iJournals All Rights Reserved www.ijournals.in
iJournals: International Journal of Software & Hardware Research in Engineering (IJSHRE)
 ISSN-2347-4890
 Volume 9 Issue 7 July 2021

 Second, they will often show decoding difficulties, that the literacy rate of children is not hampered.
 meaning that their attempts to identify words they This project aims to use rapid and holistic machine-
 do not know will produce many errors. They will learning-dependent assessment with parameters of
 not be very accurate in using letter-sound
 If one effectively processes the input parameters of
 relationships in combination with context to
 both READING (fixation/px·ms-1& saccade/px·ms-
 identify unknown words. These problems in word
 1
 ) and SPEECH TESTS (reaction time/s, reading
 recognition are due to an underlying deficit in the
 time/s, backtracking/Boolean &
 sound component of language that makes it
 pronunciation/Boolean) of a young child, THEN it
 difficult for readers to connect letters and sounds in
 is possible to quantifiably measure the degree and
 order to decode. People with dyslexia often have
 level of dyslexia using machine-learning
 trouble comprehending what they read because of
 algorithms. Our solution uses eyeball tracking and
 the great difficulty they experience in accessing the
 phonological testing because the combination of
 printed words.
 the two provides a holistic and more intricate
 4. EARLY AGE DETECTION DYSLEXIA analysis of dyslexia by improving the data set and
 accuracy of predictions.
 For many years, the importance of early
 identification and intervention for children with
 dyslexia has been stressed. Accordingly, much
 research has been directed towards establishing
 precursors of dyslexia in the preschool years in
 international studies of children at family risk of
 reading problems.

 This is the major reason being our problem
 statement was taken into consideration, as per
 reports Once children begin to think reading is
 difficult, they develop a resistance to it and lose
 confidence. Developing this attitude toward
 reading during the formative years makes it
 difficult to reverse later. To prevent this, teachers
 need to reach these students in kindergarten and Figure 2: Block Diagram for Eye Ball Detection,
 first grade. Studies show that there is little to no Explanation, and final output,
 difference between the reading abilities of average
 children and dyslexic children when intervention And this test is computer based and it takes 20
 occurs early. minutes to complete and give you the results of
 dyslexia.
 5. INNOVATION

 We are designing an objective and quantifiable
 screening of dyslexia for children at an early age so

 6
© 2021, iJournals All Rights Reserved www.ijournals.in Page 6

© 2020, iJournals All Rights Reserved www.ijournals.in
iJournals: International Journal of Software & Hardware Research in Engineering (IJSHRE)
 ISSN-2347-4890
 Volume 9 Issue 7 July 2021

 helps to track the eye while reading. Video Stream
 (imutils.video.VideoStream()) was used to give a
 live stream of the video without any break of the
 loop. Dlib is a toolkit containing machine learning
 algorithms and tools for creating complex software
 such as for this algorithm. The SpeechRecognition
 library was used to convert speech to text for the
 speaking test.

 Figure3: The flow of the screening dyslexia process

 6. ALGORITHM

 All aspects of the algorithm rely on image
 processing and speech to text software. This is
 facilitated by Python and its plenitude of libraries
 and easily-accessible server. The data collection
 happens through the connected camera and
 microphone.

 Imageio library, which provides a platform to read
 and write image data, was used to locate pupils of 7. METHODS

 the eyes. OpenCV Library, which is utilized for
 image processing, was used to trace eye Our solution works on 2 main measures: Eye

 coordinates and perform the motion tracking. tracking & Phonology.

 Moreover, Numpy (array and numerical matrices)
 library was used to convert the vision pupils to The Image Processing & Eye Land marking

 array coordinates that the programming language algorithm captures and locks the pupil coordinates

 can understand. Matplotlib, which functions on an on the screen by using the python library Imageio

 object-oriented API, was used for face plotting. that provides an easy interface to read and write a

 SciPy optimizes curve fit machine learning that wide range of image data. These coordinates are

 7
© 2021, iJournals All Rights Reserved www.ijournals.in Page 7

© 2020, iJournals All Rights Reserved www.ijournals.in
iJournals: International Journal of Software & Hardware Research in Engineering (IJSHRE)
 ISSN-2347-4890
 Volume 9 Issue 7 July 2021

 like a 2-dimensional x-axis and y-axis points on
 screen that can be pin-pointed. These coordinates
 are converted into a text file for the eyeball
 tracking mechanism, and are used with the dlib
 module. OpenCV library and its classifiers gauge
 distance between eyes and the border of the face.
 Thus track the eyes movement specifically using
 CV2 by creating a mask using matplotlib.

 The fixations and saccades are recorded by the
 motion of the eyes while reading. Bresenham's
 algorithm, which works on the principle of
 incremental scan conversion, relies only on integer
 calculations for accurate tracking of eye movement.
 The curve fit function of the SciPy library is RESULTS AND CONCLUSIONS
 utilized to create a mathematical function to model
 the data points into text which are then statistically After conducting numerous trials and consistently
 compared to determine the extent to which the improving the efficacy of the machine learning
 child is dyslexic. program, it was found that both the reading and
 speaking tests were possible and provided valuable
 Speech recognition software uses py-audio and information about the levels of dyslexia. The
 pyttsx3 libraries to capture the speech of the user reading test, when taken by dyslexic children, was
 and identify patterns or characteristics commonly extremely effective in identifying them as such
 associated with dyslexia such as backtracking and with an accuracy as high as 86%. The speaking test
 grapheme mispronunciations. Furthermore, we did have a few errors in the program running at
 incorporated additional filtering processes of font first. But once they were resolved, it functioned
 and size of texts which were specific to age well, although the accuracy of the algorithm needs
 categories. Thus, the reading and speaking tests to be improved.
 will furnish customized feedback to the child.
 The combined run of the two tests led to 91%
 accuracy with SVM model and 90.9% accuracy
 with logistic regression. The speaking test did have
 a few errors in the program running at first. But
 once resolved, it functioned well. Compared to
 literature values, results are lower but very
 competitive. With expanded datasets and
 increasing accuracy in feature extraction
 algorithms, this can be surmounted

 8
© 2021, iJournals All Rights Reserved www.ijournals.in Page 8

© 2020, iJournals All Rights Reserved www.ijournals.in
iJournals: International Journal of Software & Hardware Research in Engineering (IJSHRE)
 ISSN-2347-4890
 Volume 9 Issue 7 July 2021

 and accessible to the young generations of our
 country.
 9. ACKNOWLEDGEMENTS

 This research was supported by On My Own
 Technology Pvt Ltd. (OMOTEC) We thank our
 mentors who provided insight and expertise that
 greatly assisted the research, although they may not

 Moreover, the tests were effective for each age- agree with all of the interpretations/conclusions of

 category as they used passages of varying fonts, this paper.

 sizes and spaces to allow for the reading ability of
 that age. The OpenCV library and Bresenham's
 algorithm significantly improved the accuracy of
 the results, however larger and more extensive data
 sets are required to further improve the accuracy of
 the algorithm. Screener that effectively screens for
 early literacy milestones- tablet based, self
 administered, scientifically validated and ready for
 scaling. Data analytic for customized resources
 support for coaching- data back to solutions

 ● Distribution- schools need for test scores,
 screening legislation
 ● AI technologists
 REFERENCES
 8. FUTURE SCOPE [1]"Dyslexia FAQ." Yale Dyslexia. Accessed
 A major strength of this prototype is that it is November 13, 2020.
 highly standardized in equipment yet personalized https://dyslexia.yale.edu/dyslexia/dyslexia-faq/.
 in feedback to every child. This ensures that the
 prototype can be scaled to numerous schools and [2]Casalis, S. (2004). The concept of dyslexia. In
 platforms through state associations such as the Handbook of children’s literacy (pp. 25’ure, and
 Maharashtra Dyslexia Association and national prospects. In Proceedings of the world congress on
 associations such as Dyslexia Association of India. dyslexia on reading, perception, and language (Vol.
 A factor that must be improved is its accuracy and 12, p. 272)
 potential economic feasibility for rural schools but
 with persistent evaluation and detailed procedures, [3]Desroches, Amy S., Marc F. Joanisse, and Erin
 this too can be accomplished. K. Robertson. "Specific Phonological Impairments
 Our prototype-the newest and most effective in Dyslexia Revealed by Eyetracking." Cognition
 assessment method for dyslexia- will be scalable 100, no. 3 (2006).
 doi:10.1016/j.cognition.2005.09.001.

 9
© 2021, iJournals All Rights Reserved www.ijournals.in Page 9

© 2020, iJournals All Rights Reserved www.ijournals.in
iJournals: International Journal of Software & Hardware Research in Engineering (IJSHRE)
 ISSN-2347-4890
 Volume 9 Issue 7 July 2021

 AUTHOR PROFILE
 [4]Jones, Manon W., Mateo Obregón, M. Louise KareenaShankta
 Kelly, and Holly P. Branigan. "Elucidating the Completed Cambridge International General
 Component Processes Involved in Dyslexic and Certificate of Secondary Education from Dhirubhai
 Non-dyslexic Reading Fluency: An Eye-tracking Ambani International School Currently studying
 Study." Cognition 109, no. 3 (2008): 389-407. for the International Baccalaureate Diploma
 doi:10.1016/j.cognition.2008.10.005. Programme of 2022 Mumbai, India
 Email ID: kshankta|@icloud.com
 [5]Mahmoodi-Shahrebabaki, M. (2014). E-learning
 in Iran as a developing country: challenges ahead Reetu Jain
 and possible solutions.International Journal of Chief-Mentor and Co-founder of On My
 Research in EducationMethodology,6(2), 788-795 OwnTechnology Private Limited Mumbai,
 Email:reetu.jain@onmyowntechnology.com
 [6]Mahmoodi-Shahrebabaki, M. (2018, February).
 Dyslexia: past, present, and future. Paperpresented Shekhar Jain
 at 6th Annual MTSU Literacy Research Chief Executive Officer and Co-founder of On My
 Conference, Murfreesboro, TN. OwnTechnology Private Limited Mumbai,
 Email:shekhar.jain@onmyowntechnology.com
 [7]NilssonBenfatto M, qvistSeimyr G, Ygge J,
 Pansell T, Rydberg A, Jacobson C (2016) Radhakrishna Madas
 Screening for Dyslexia Using Eye Tracking during Sr. Research Associate at On My Own Technology
 Reading. PLoS ONE 11(12): e0165508 Private Limited Mumbai,
 Email:radhakrishna.madas@onmyowntechnology.c
 om

 10
© 2021, iJournals All Rights Reserved www.ijournals.in Page 10

© 2020, iJournals All Rights Reserved www.ijournals.in
You can also read